The 1-bit revolution, GPT-Live, Apple pains, & Lots of Models

TLDR;

Welcome back! Today, we have several new models and products, new maths discoveries from OpenAI, market troubles, the fascinating concept of 1-bit models, GPT-Live’s unique architecture, and more.

Enjoy!

FRONTIER RESEARCH

OpenAI’s Astra 10 new math discoveries

As published by OpenAI, an internal version of its upcoming Astra model generated solutions or substantial progress on ten longstanding problems in mathematics and theoretical computer science.

The problems span geometry, coding theory, group theory, quantum computing, cryptography and combinatorics; the results include improved mathematical bounds, proofs of computational hardness, the resolution of several Erdős problems and counterexamples disproving major conjectures.

OpenAI said the model produced the underlying arguments, while humans helped turn them into research manuscripts. The model then translated every proof into Lean, a system used to formally check mathematical logic.

OpenAI estimates that generating all ten solutions required about $2,000 worth of computation at its Sol API pricing, and says it is taking responsibility for the correctness of the published results while inviting further scrutiny from mathematicians.

TheWhiteBox’s takeaway:

Very impressive work from OpenAI. But just like I mention below with Qwen’s new model, or every new “discovery” we see from models, it’s important that we’re able to constrain how impressed you should actually be, because these discoveries are all happening in science domains for a reason.

When a task is mathematically verifiable, meaning a machine can produce outcomes and easily verify whether they are right or wrong, there’s really no upper bound on what AI can eventually achieve, insofar as it has enough compute.

If you can represent a problem as a mathematical function, AI is nothing but an optimization over that function, turning the process into an exploration of the best possible solution.

These results are, of course, transformational and cause for celebration because there are multiple areas where verifiability is a given, like maths or coding. However, success isn’t remotely as clear when the task is not verifiable, like writing a book.

AI can ‘hill climb’ itself into becoming better at maths by “simply” finding the best solutions and learning them. For lack of a better term, it can ‘brute force’ its way to better solutions; it’s just a matter of experimentation, time, and compute. But climbing that hill in areas where the model can’t easily correct (or worse, improve) is much more complicated. Or outright impossible.

There’s a reason why we don’t have superhuman AI writers. Just like you don’t have a mathematical formula to output New York Times best sellers, AI doesn’t have a clear-cut way to find the “solution” to producing NYT best-selling books.

SMALL MODELS

Running Great Models Locally at 240 tokens/second

I wrote an entire Medium article for this model in case you want to dive deeper; you can find it here.

The extremely-low precision wave is unstoppable. And the next evidence of this is coming from a startup called DeepGrove and their ternary-weights model called Maple.

This model, which would have been the state-of-the-art of AI back in September 2024,…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.